Application of ComBat-seq on simulated datasets

Levels of dispersion differences and confounding

About this simulation

From previous analysis (“powerdown_unbalanced”), it appears that ComBat-seq works when there is variance batch effect AND when there is some level of confounding between batch and condition (i.e. in unbalanced study design). We dig deeper in these two factors.

Simulation setup

• Mean fold changes: Biological - 2, Batch - 1.5
• Dispersion: Batch 1 - 0.15, Batch 2 - 0.15*dispersion fold level
• Study design: Different levels of confounding
• Reads per gene: Median 120 (coverage 5)
• Total number of samples: 20 (we look at N=60 later in comparison to see if patterns hold)

Note that the fixed parameters are consistent with (i.e. within the range of) my observations from real datasets (See explore_real_data for details).

Two factors under examination are dispersion fold change and level of confounding. Confounding level is measured by how many percentage of samples in batch 1 is control samples. 2 batches have the same size, but in batch 2, the number of controls and cases are reversed compared to batch 1. Grids for the two factors:

• Dispersion level: 1 (no dispersion difference), 2, 3, 4, 5
• Confounding: 0.5 (no confounding), 0.4, 0.3, 0.2

Results

TPR vs FPR

Line plots

Facet by confounding levels

Facet by dispersion levels

Scatter plots

TPR (sensitivity) vs FDR (1-precision)

Line plots

Facet by confounding levels

Facet by dispersion levels

Scatter plots

Similar plot (sensitivity vs 1-FDR) as Figure 6 in DESeq2 paper.

Thoughts & future work